Research On Deep Multi-view Clustering Algorithm Based On Graph Embedding

As a common unsupervised learning technique,clustering aims to partition samples in a dataset into distinct classes,such that samples within the same class exhibit similar features while features vary significantly across different classes.Currently,clustering techniques have become ubiquitous in data mining,machine learning,pattern recognition,and related fields.Previous clustering algorithms such as k-means can effectively handle data with a single feature.However,with the advent of multi-view data,these single-view clustering methods have faced significant challenges.In order to address the multi-view data in a scientifically rigorous manner,multi-view clustering has emerged.Since multi-view clustering takes into account the consistency and complementarity between different views,it is able to synthesize the information of multiple views and improve the accuracy and stability of clustering,which has become one of the main research directions at present.How to maximize the consistent and complementary information between multi-views to generate a better clustering structure has become a pressing problem to be solved.Traditional multi-view clustering methods usually try to find a common latent space between multi-view data and then employ techniques such as k-means or spectral clustering to obtain clustering results.These methods are time-consuming and labor-intensive while separating feature extraction and clustering,which cannot be learned jointly.Recently,the development of deep learning has brought new ideas for multi-view clustering,and combining multi-view clustering and deep learning has become a hot research topic nowadays.However,the existing deep multi-view clustering methods rarely consider the correlation information between data,resulting in limited improvement of the clustering performance.In order to solve the above problems,this thesis proposes a deep multi-view clustering framework based on graph embedding to improve the deep clustering model by fusing the graph information of the original data to impose constraints on the clustering layers.The main research work of the thesis is as follows:Firstly,multiple autoencoders are used to explore consistent and complementary information from multi-views and find the common latent space simultaneously.In addition,a new graph embedding scheme is proposed considering the validity of the nearest neighbor correlation and local structure information of the original data.Specifically,the graph affinity information of the original data is directly applied to the soft assignment of the data,consistent with the clustering loss,to improve the performance of multi-view clustering.Experimental results on several datasets verify the effectiveness of the method.Secondly,since graph embedding terms are very dependent on the initial graph affinity information,different affinity matrix construction methods have a great impact on the clustering results.Besides,it is difficult to guarantee the full accuracy of data affinity by any method of constructing affinity matrix,and the original affinity relationship between data may change with the update of clustering process.To solve this problem,this thesis designs an optimization method that can dynamically adjust the data affinity.The information entropy is introduced to adjust the neighbor relationship of the data,so as to better guide the clustering.The experimental results show that the information entropy constrained nearest neighbor affinity improvement algorithm exhibits a superior clustering performance.Finally,the two previously proposed clustering methods employ KL scatter for clustering,which enhances the soft distribution of data.However,KL scatter lacks sufficient discriminative power to distinguish between different cluster data distributions,often resulting in samples from similar but distinct classes being grouped together in the same cluster.This limitation makes the methods vulnerable to interference and lacks robustness.Therefore,this thesis introduces an adversarial learning mechanism,which uses discriminative networks to classify samples into classes instead of KL scatter.The experimental results show that the adversarial learning mechanism can further improve the classification ability of the model,thereby enhancing the clustering effect.